Decision fusion using fuzzy integral method
Jimin Liang, Wanhai Yang, Xiyao Cai · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2000
It is an intuitionistic demand for multisensor decision fusion processes that the degree of importance of each sensor's local decision towards the final decision should reflect adequately its detection, identification or classification ability, especially in time variant environments. This can be seen clearly in distributed detection systems, where the optimal final decision is the weighed sum of the local decisions, and the weights are functions of the local sensor's detection and false alarm probabilities. When voting fusion strategy is used, it supposes in nature that all the sensors are working at the same performance level. When dealing with soft decision fusion problems, the local sensor's reports represent its uncertainty about the observed phenomena. The most commonly used approaches for soft decision fusion are Bayesian probability theory and Dempster-Shafer evidence theory. Both of them update the evidence using the sensors' reports only, the reliability differences between the decision-making sensors are not considered. In this paper, the decision fusion ability of fuzzy integral is investigated from the viewpoints of both hard decisions and soft decisions. When fuzzy integral is used to fuse multisensor decisions, the fuzzy density can be interpreted as the importance of each sensor's local decision towards the final decision. Fuzzy densities can be subjectively assigned by experts, or be induced form data set. We present a learning algorithm to distinguish the degree of importance of each sensor's local decision. A genetic algorithm-based hybrid learning method is used to determine the fuzzy densities from the training data set. A multi-pattern classification experiment is simulated to validate our proposed method.